acceptodds
Under review as a conference paper at ICLR 2027

Query-Count-Independent Differentially Private Synopses for Retrieval-Augmented Generation

Abstract

Retrieval-augmented generation (RAG) can expose protected records through repeated access to its knowledge base. Query-time differential privacy (DP) requires cumulative accounting, which can constrain utility under a fixed total budget. Reusable DP synthetic corpora avoid per-query expenditure but face the challenge of preserving evidence for downstream retrieval. We propose SynopsisRAG, an offline-release framework that uses protected evidence to guide the private selection of intact public passages. A bounded fidelity score combines assessed answer completeness with estimated private-evidence support, guiding bidirectional reweighting and privacy-calibrated sampling of one representative per category. Online retrieval and generation then reuse the released synopsis without accessing the private corpus or incurring additional privacy loss for that version. We establish pure chunk-level DP with a direct document-level extension, derive conditional answer-support bounds for the released synopsis, and quantify how privacy calibration limits release concentration. On healthcare and finance workloads of up to 10,000 queries, SynopsisRAG achieves higher long-horizon match accuracy than the evaluated query-time and synthetic-corpus DP baselines. Extraction tests additionally show reductions in measured token leakage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation, Synopsis, Differential Privacy

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BhXmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1(−1)
Soundness
2
Read review
Paper summary

The paper proposes a DP framework that eliminates per-query privacy expenditure by privately selecting a fixed set of public passages during an offline preparation stage. It combines answer completeness and private-evidence support into a bounded fidelity score, then applies bidirectional multiplicative-weight updates to prioritize informative passages and underrepresented queries. The authors establish pure DP guarantees and conditional answer-support bounds.

Critical strengths
  1. The offline synopsis construction decouples privacy expenditure from query volume, addressing a limitation of query-time DP mechanisms in long-lived RAG systems. Selecting intact public passages also preserves source provenance while avoiding direct exposure of protected retrieval content.
  2. The paper establishes pure DP guarantees for the sampled synopsis, including the effects of data-dependent reweighting, and extends these guarantees to adaptive query transcripts and document-level adjacency. The analysis also explicitly distinguishes estimated evidence support from semantic correctness, making the theoretical assumptions relatively transparent.
  3. Experiments on healthcare and finance datasets demonstrate improved long-horizon accuracy over several DP-RAG baselines, with reduced empirical extraction leakage.
Critical weaknesses
  1. The framework combines offline DP release, multiplicative-weight updates, and evidence-support estimation. The paper does not adequately isolate the contribution of bidirectional reweighting, private-evidence guidance, and the proposed fidelity score. Comparisons against uniform sampling alone are insufficient to establish the necessity of the individual components, particularly given the existing related work on private synopsis construction.
  2. SynopsisRAG releases only one public passage per category, making retrieval capacity limited. Its effectiveness depends on public passages covering information present in the protected corpus and on evaluation queries aligning with predefined categories. Consequently, the method could not recover private-only knowledge and may struggle with heterogeneous, compositional, or multi-document questions. The evaluation does not examine open-domain routing or systematically quantify performance when relevant information is absent from the public corpus.
  3. Although the paper proves bounds on low-support risk, its answer-validity guarantee includes an unmeasured semantic-error term, meaning it does not directly establish answer correctness. Moreover, the privacy calibration constrains selection concentration. This raises questions about how much meaningful private-evidence-driven selection is possible under strong privacy constraints. The experiments do not adequately characterize this relationship or report uncertainty across independently sampled synopses.
Questions for authors

How to address the concerns mentioned in the Weaknesses section.

Reviewer_geSpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SynopsisRAG makes RAG private by using the private corpus only offline, to score and select one intact public passage per topic, released through an ε-DP exponential-weights sample. All later queries are answered from these released public passages alone, so they cost no extra privacy budget, and the authors report higher long-horizon accuracy than query-time and synthetic-corpus DP baselines on healthcare and finance.

Critical strengths

The paper is well written and easy to follow, and it addresses an important domain: privacy for long-lived RAG services over sensitive knowledge bases.

Critical weaknesses

The method protects privacy mainly by not using the private data. The generator never sees private text, and the only private-data-dependent output is the choice of one public Wikipedia passage per category. No private facts can reach the answer, which sidesteps the main goal of private RAG. Uniform release and public-domain RAG already give perfect privacy and are within a few points of SynopsisRAG (50% vs 54% at ε=1). For the same reason, the token-leakage comparison with DPRAG and MURAG is not like for like.

The private signal barely affects the release. By the paper's own Proposition 2, at ε=1 and K=10 any two candidates differ in probability by at most about 1.05×, so the fitted distribution is essentially uniform. Even at ε=10 the ratio is only about 1.65×. The gains over uniform release are therefore hard to attribute to fitting rather than to single-draw luck, and no variance across repeated synopsis draws is reported.

The design rests on many unjustified empirical choices. These include K=10 clusters, 32 candidates and 5 seeds per category, FLAN-T5 coverage labels, MiniCheck support, a hand-written applicability predicate, λ=0.5, k=3, τ=0.7 and μ=0.4. Only ε, K and k are ablated, and accuracy drops from 74% at k=1 to 61% at k=3, suggesting that more private evidence hurts. The theoretical bounds depend on the semantic error term ξ_c, which is never measured.

The models are small and dated. FLAN-T5-XL and MiniCheck-Flan-T5-Large, both with 512-token caps, drive every score, and their accuracy is never evaluated. The generators are LLaMA-2-7B and Mistral-7B from 2023. Results with current models such as Qwen2.5 or Llama-3 would be more convincing.

Related inference-time DP work is missing. PMixED (Flemings et al., NAACL 2024) mixes a private next-token distribution with a public one under a Rényi divergence bound. In RAG, the context with and without retrieved private documents can be mixed the same way, as done in DP-Fusion (arXiv:2507.04531). None are cited or compared against, even though these methods let private evidence reach the answer with bounded influence.

Questions for authors
  1. What exactly does the method protect in practice? Since the generator only ever sees public passages and no private facts can reach the answer, what practical benefit does the private corpus give over public-domain RAG or uniform release?
  2. Why were FLAN-T5-XL and MiniCheck-Flan-T5-Large chosen for scoring, and LLaMA-2-7B and Mistral-7B for generation? How do the results change with current instruction-tuned models such as Qwen2.5 or Llama-3? How accurate are the selector and checker judgments?
  3. Why not mix private and public contexts at inference time, as in PMixED (Flemings et al., NAACL 2024) and DP-Fusion (arXiv:2507.04531)? Running the LLM with and without the retrieved private documents and mixing the two distributions under a divergence bound would let private evidence reach the answer. How does SynopsisRAG compare against such methods at matched privacy budgets and query counts?
  4. How were the topics, the number of categories, candidates and seed questions, and the thresholds τ, μ, λ and k chosen? How sensitive are the results to these empirical choices?
Reviewer_zU5oOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SynopsisRAG targets long-lived RAG services over a protected corpus. Instead of spending privacy budget per query, it releases once, offline, one intact public passage per category (default K = 10 categories, 32 public candidates each), and answers every query from the released passage of its routed category, so all online interaction is post-processing. Candidates are scored by multiplying a public completeness grade (FLAN-T5-XL) by MiniCheck support from the top-k retrieved private windows; scores are clipped to [-1, 1], passed through t = 2 rounds of multiplicative-weights updates over candidates and seed queries, and one candidate per category is sampled with step size ε_c/(4t). The paper proves pure ε-DP at chunk and document level, gives average and population low-support bounds and a concentration bound (Prop. 2), and reports higher match accuracy than DPRAG, MURAG and DP-SynRAG over up to 10,000 queries on healthcare and finance with LLaMA-2-7B and Mistral-7B (51% vs 46/30/18% at ε = 1 on healthcare), plus lower token overlap under four extraction attacks.

Critical strengths
  1. A clean deployment pattern with a correct privacy proof. Private data is touched only offline, the release is a single sample over public passages, and serving is post-processing, so total privacy loss is fixed regardless of query count. The proof correctly covers the data dependence of both fitting distributions. This is the main reason the approach deserves attention for long-lived services.
  2. Intact public passages with provenance. The release contains no fabricated synthetic text, and the paper specifies precisely what is published and what stays with the curator, which matters in regulated domains such as healthcare and finance.
Critical weaknesses
  1. At the evaluated budgets, the paper's own Prop. 2 caps the gain from private fitting below the reported gains. Prop. 2 keeps every release probability within a factor e^{±ε_c/2} of uniform for any private corpus: with the defaults (K = 10, ε = 1, 32 candidates), each candidate is drawn with probability 0.0298 to 0.0328, against 0.03125 under uniform release. The expected accuracy gain over uniform release is therefore at most tanh(ε_c/8) ≈ ε/(8K): 0.6, 1.25, 6.2 and 12.4 points for ε = 0.5, 1, 5 and 10. The reported gains over uniform release (50% at 500 queries) are 1, 4, 11 and 17 points (Table 2, matching Fig. 5 at 500 queries), and 24 points for k = 1 at ε = 5, where 74% also exceeds Pure RAG (70%). Accuracy stays at 58-61% as K grows from 5 to 20 and the per-category budget shrinks 4×. Neither the number of release draws nor any variance is reported, so these differences reflect draw-to-draw variance or an implementation that departs from the analyzed mechanism. The cause is structural: one private chunk can move f from 0 to 1 for every (q, a) pair, so the score's sensitivity equals its range and corpus size cannot sharpen the release, which at ε = 1 is the exponential mechanism over mean_q s_{q,a} up to a 2.5% query-weight perturbation (McSherry & Talwar, 2007, uncited). The central claim, that private evidence improves the release, is therefore untested.
  2. The evaluation omits the free baselines that would test private guidance, and its leakage metric cannot isolate leakage. Sec. 3.1 calls for a public-only selection with the same candidate space and capacity; argmax_a mean_q b(q, a) is the natural one, since b is computed from public data (Table 1; Alg. 1, line 17), yet the experiments use uniform release. At 500 queries (Fig. 5a), Pure LLM scores 48%, public-domain RAG 48% and uniform release 50%, and none of these free baselines appears in the 10,000-query plots, where DPRAG and MURAG hold at 15-18% and 30% from about 100 queries onward (Fig. 2a). A fallback to no-retrieval generation would hold about 48%, so the long-horizon gap measures budget exhaustion without fallback. The like-for-like baseline, DP-SynRAG, trails by 5 points, and the shaded bands that would show whether this exceeds run-to-run variation are undefined. The leakage metric charges SynopsisRAG, whose generator never sees private text, 375 tokens under Pirate against 399 for DPRAG, so the 24-token gap cannot be attributed to private leakage; the 55.6% headline compares 4 with 9 tokens, and DP-SynRAG also leaks 4. Neither the utility nor the leakage advantage is established.
Questions for authors

Clarification requests:

  1. How many release draws underlie each number? Please report mean ± std over at least 20 draws for SynopsisRAG and uniform release and the realized max_a P(a) per category, and confirm that the experiments use α_s = α_q = ε/(4Kt) with a categorical sample rather than an argmax.
  2. What are the 10k-query accuracies of Pure LLM, public-domain RAG, and the deterministic argmax of mean_q b(q, a)? What fraction of (q, a) pairs have s > -1?
  3. For DPRAG, MURAG and DP-SynRAG, what are the per-query ε, the behavior after budget exhaustion, the adjacency and δ?
  4. How is the evaluation stream sampled and ordered, is accuracy at n cumulative, are seed questions deduplicated against evaluation questions, and what do the shaded bands in Figs. 2-3 show? Why does the fixed SynopsisRAG synopsis keep losing accuracy after 100 queries while Pure RAG stays flat (Fig. 2a)?

Suggestions for future work:

  1. Aggregate private evidence with bounded per-record votes (DP nearest-neighbor histograms as in Private Evolution, Lin et al., ICLR 2024, or report-noisy-max over support counts), so sensitivity is 1 while counts grow with the corpus.
  2. Estimate W(q, a) and ξ_c(P) with independent annotation to connect Prop. 1 to the experiments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.